Content uploaded by Eduardo Feo Flushing
Author content
All content in this area was uploaded by Eduardo Feo Flushing on Jan 03, 2023
Content may be subject to copyright.
Prediction of Indoor Wireless Coverage from 3D
Floor Plans Using Deep Convolutional Neural
Networks
Yaqoob Ansari, Nouha Tiyal, Eduardo Feo-Flushing, Saquib Razak
Department of Computer Science, Carnegie Mellon University in Qatar, Doha, Qatar.
Email: {yansari,ntiyal,efeoflus,srazak}@andrew.cmu.edu
Abstract—With significant advancements in Machine Learning
and Deep Learning, Convolutional Neural Networks (CNNs) have
shown promising results in handling classification and regression
problems. We present 3DRSSNet, a model that can predict the
signal strength of an Access point in a 3-D environment based
on the 3-D floor map of the building. Our deep CNN approach
differs from previous work in that (i) it can generalize to unseen
environments, and (ii) to the best of our knowledge, it is the first
work to utilize 3D maps to build a signal strength prediction
model and validate its results using actual measured data.
The proposed neural network model can help solve problems
lie optimal access point placement and blind spot detection.
Experimental results show that CNNs can predict indoor radio
link quality with a performance with a Mean Absolute Error of
4.4 dBm and is able to generalize well to unseen environments.
Index Terms—Deep Learning, Radio Link Quality, RSS, In-
door Path Loss, Data Packets, Floor Plans, Neural Networks,
CNN, Regression,
I. INTRODUCTION
Ensuring a stable and reliable platform for communication is
one of a wireless network’s most crucial roles. The network’s
ability to provide this platform for use is directly linked to
the estimation of its radio link quality. The radio link quality
estimation in wireless networks plays a fundamental role in
designing its higher-layered protocols and directly relates to
the network’s performance. Learning ways to estimate and
optimize the link quality and make correct spatial predictions
of the value is vital for some wireless network protocols and
applications.
In wireless networks, link quality is generally estimated by
measuring Received Signal Strength (RSS). Several radio wave
propagation phenomena such as diffraction, scattering, trans-
mission, refraction, and reflection, have an effect on the RSS.
Furthermore, in the majority of circumstances, the transmitter
and the receiver are not in direct line-of-sight As a result, the
RSS varies in time and location with regard to the receiver
and transmitter [1]. The complex interplay among all these
factors make precise link quality prediction a challenging task
in wireless networks. This problem is exacerbated in indoor
wireless networks because the involvement of the environment
(e.g., walls, ceilings) in determining signal strength at a given
place is difficult to compute [2], [3].
Given the growing importance of predicting the RSS val-
ues and the task’s complexity to take into account multiple
variables, we propose, 3DRSSNet, a novel neural network
architecture for indoor radio link quality estimation by pre-
dicting the Received Signal Strength (RSS) value that dif-
fers in two critical aspects compared to most existing DL
approaches. Firstly, it considers the complete 3D structure
of the surrounding, accounting for all the objects and other
obstacles, predicting scattering and blockage of the signal by
the obstacles. Secondly, it predicts the value independently
of the location of the access point and the receiver. The
model is built on the idea that the surrounding environmental
elements play a vital role in determining the signal strength at
a location. It uses a regressive Convolution Neural Network
(CNN) approach that responds very well to the importance of
adjacent measurement in a 3D topology. CNNs are considered
”universal function approximators”, which allows them to
learn the relationships between 3D objects, signal strength
measurements, and the signal strength field [4].
II. RE LATE D WORK
Existing related works on radio link quality prediction
in general differ in the assumptions made, the scenarios of
experimentation, and the methods employed. It was noted that
the results produced were radically different, sometimes even
contradictory. Another difference is the metric monitored, e.g.,
RSS [5], Packet Reception Rate (PRR) [6], among others.
Different approaches have been used to predict wireless link
quality. Some works aim to create mathematical models to ap-
proximate the physical aspects of the wireless propagation [7].
These models rely on numeric data that is backed by physics
as their basis for model creation. Other works aim to create
models using intelligent algorithms such as machine learning
and deep learning tools to learn features and estimate the link
quality [8], [1], [9], [10], [11]. These works are more data-
driven in the context of information and aim at detecting and
learning from features of the provided data.
The selection of data to train the models and the prediction
target can play a crucial role in determining the performance
and accuracy of the model. From attempted known works,
it becomes apparent that models which rely on classification
tend to underperform as they cannot predict the values but
only classify them over ranges. Regression-based models are
trained with numeric data from simulations (e.g., [11]), but and
also tend to underperform as they fail to consider the impact
of the surrounding environment of the wireless network. This
raises the need for a regressive model with considerations for
environmental and other numeric factors.
Easy accessibility to 2D environmental representations (like
satellite images [12], and city-maps [11]) prompts most works
to rely on them as inputs. However, the use of 2D data can
significantly limit the information about a structure. Other
approaches combine interpolation techniques with machine
learning models to estimate radio spectrum metrics [10].
These approaches require data to be collected from the same
environment where the predictions are made.
We propose a novel approach to learn the factors that affect
radio link quality in wireless signals. Our research attempts to
address some of the previously stated omissions and presents
a new approach to predict link quality by recording the RSS
value while considering the multilayered floor plan of the
structure. We let the model detect and understand features
more explicitly by providing the 3D floor plan with three
information channels.
With such assisted training, the model can learn to predict
the signal strength with a manageable loss factor and develop
practical generalization capabilities.
III. PREDICTION MOD EL
We use a deep learning model to capture the relationship
between the 3D map of the AP’s environment and the specific
receiver point and the expected signal strength at a location
in the volume-of-interest (VoI). The model can be trained
beforehand, using data collected from other environments. The
space is discretized into voxels (3D volume ”pixels”), where
each voxel is of dimension δx×δy×δz. The aggregate of all
voxels is a tessellation of the 3D Euclidean space. The set of
all voxels that are part of the VoI is denoted as V. The location
of a sender vsand receiver vrare represented as voxels,
vs, vr∈ V. Each voxel is associated with a tuple value whose
first component indicates the environment elements contained
in it, e.g., walls, furniture, and free space, while the second
component indicates the relevance of the voxel to the signal
propagation from vsto vr.
Under these considerations, a measurement is characterized
by a VoI (i.e., the set of voxels V), the location of the AP (vs),
and the location of the receiver (vr). By assigning a special
numeric value to vsand vr, all the elements that describe a
measurement can be conveniently encoded in a 4-dimensional
tensor, denoted as x. The signal strength value, in dBm, at vr
is denoted as y∈R.
Our goal is to cast the prediction of the RSS at a given
location as a regression problem, where we learn a function
y=f(x), such that, given a VoI and the location of the AP
and the receiver, encoded as x, provides an accurate prediction
of the RSS y.
IV. DATAS ET G EN ER ATIO N
In this section, we describe our data collection methodology.
A 3D floor plan is translated into a 3D model following a
manual procedure. The 3D models are then voxelized, where
corresponding voxels are selected from where RSS measure-
ments are taken. For each sampling location, a 4-dimensional
tensor is constructed by assigning numerical values according
to the location of the sampling point. In the end, we obtain a
complete data sample, which is added to the set of all samples.
A. 3D Floor Plan creation
A 3D model design software was used to create the 3D
floor plan. The 2D floor plans were colored according to
the material of the objects in it. This color encoding was
consistent throughout the dataset. The 3D models were stored
in a file format that could retain the different color and texture
properties that were imparted to distinguish the objects from
one another.
B. Voxel Representation
We used an automated tool to voxelize a 3D model. Given
the voxel dimensions, the software returns a list of (x, y, z, r)
values, where (x, y, z)is the location of the voxel in the
euclidean space, and ris an integer value (i.e., the color)
associated to the voxel. The software was also used to fine-
tune the coloring of the voxels and to add the location of the
access point.
C. RSS samplings
The 3D floor plan was systematically divided into location
points for data collection. The Location points were 30 cm
apart from each other in the xy-plane and were set on three
different zor height values. The voxel editor tool was used to
determine the coordinates of the sampling locations. At every
location, the RSS value was estimated for a duration of 3
minutes.
Measurement of RSS was made at 2.442 GHz carrier-
frequency using an AC1750 DB Wireless Dual-Band AC+
Gigabit Router with Tx signal power 3 dBm and bandwidth 20
MHz . Network traffic was generated from a second device in
the wireless network to guarantee a continuous flow of packets
from the AP. The mean of the data collected was considered
as the final value for each location.
D. Matrix creation
Once RSS estimates from all points in VPare defined, we
proceed to the generation of the dataset. For each point in vr∈
VP, we define a tensor xof fixed dimensions L×W×H×2.
Any voxel not in the VoI is assigned values (0,0). Voxels
in the VoI are assigned values (a, b), where ais the color
assigned to vrand value bwas assigned based on distance; 0
denoted distances over 20 meters, 10 denoted distance under
5 meters and a value 5denoted the rest of locations.
E. Data Augmentation
A total of 813 location points were initially collected. This
amount of data seemed insufficient, and the early model
training with this dataset yielded considerably larger loss
values.
To augment the data, we randomly select a voxel within
1 meter distance of the original sample and then find the
corresponding locations where only few RSS values were
obtained and generate artificial samples. 1712 samples were
obtained after performing the augmentation process.
V. CNN MOD EL
A CNN model was used due to its spatial filters that
establish a ”local connectivity pattern between neurons of
adjacent layers” [4], making the architecture most suitable
for spatially local input patterns such as our 3D data. The
regression neural network model was made using Keras and
scikit-learn libraries.
A. Model Architecture
The model architecture contained a total of 9 layers. A
Convolutional Neural Networks (CNN) layer acted as the input
layer, and a Dense layer was the output layer. We used the
default parameters for Adam in Tensorflow. We used a starting
learning rate of 10−5with a batch size of 64. While training,
the model was fitted with a learning rate scheduler that altered
every epoch’s learning rate to the best rate to train the model.
The data was trained and tested on a balanced data set, with
appropriate data added to make the distribution equal. This
was done to allow the model to predict better and generalize
over a more extensive range of values. Moreover, the model
weights for the best performances were saved from training.
B. Loss Function and Model Evaluation
Mean Absolute Error (MAE) was used to track model
performance. The loss measures were put in place to check
and avoid over-fitting and under-fitting of the model. Model
regularization strategies such as adding dropout layers and
early stopping were implemented to train the model optimally
for our required task. Plots of the training and validation losses
were also generated to aid the model’s training visually. The
data was tracked on its performance on the metric of MAE,
which is the difference between the actual measured data and
the predicted data.
VI. RE SU LTS
The data set generated was used to train multiple models.
The models differed in their architectures but were trained for
an equal number of epochs and the losses were compared to
ensure optimal training loss.
Results after 200 epochs show that 3DRSSNet displayed the
best results, with its training loss staying in the range of 2-3
dBms and its validation loss in 4-5 dBms.
The training of 3DRSSNet was closely monitored, ensuring
that the model does not over-fit or under-fit. Upon extensively
testing, the model achieved a prediction error of 4.4 dBms,
as shown in both Fig. 1a and Fig. 1b, on the test data
sample, showing sufficiently accurate predictions and suitable
generalizing property.
The performance of 3DRSSNet was compared with the
performance of models of other known work in this field.
Most existing models differ in the training data used. Data can
be generated by using ray-tracing software [15] or manually
(a) Training loss and Validation loss of 3DRSSNet
(b) Model results on test data set
Fig. 1: 3DRSSNet Performance
collected at different sites (indoors or outdoors). Most models
also differ on the floorplan format. Models that utilize a 3D
representation of the floorplan and use data simulated from
ray-tracing software and a CNN architecture such as [13]
produce a model with MAE loss of 6.95-11.11 dBms. Models
that use a 3D vector representation and use manually gathered
data for training an Artificial Neural Network (ANN), which
is a promising architecture for this problem [16], [17], produce
a model with MAE loss of 15.52-16.21 dBms[1]. Models
trained without utilizing the floorplans; like the ANN trained
by [14] which uses no floorplan and manually measured
values to produce a model with MAE loss of 5.37-5.65 dBms.
Comparison of 3DRSSNet with these three models highlights
our model’s high performance with only MAE loss of 4.4-5.5
dBms and the need to use the 3D representation of floorplans
to better aid models in predicting signal strength.
The performance of 3DRSSNet was also compared with the
RSS estimations obtained from ray tracing simulations. The
ray tracing simulations predicted values with a MAE loss of
11.23 dBms. This highlights the need for CNN based model
to better learn factors that impact the signal strength.
The model was trained on data with and without data
augmentation. This was done to understand the impact of
increasing the training data’s size on the loss of the model.
The loss values were higher (in the range of 40 dBms) than
when trained with augmented data.
Author’s Work Floor Plan Format Data source Model Algorithm MAE (dBm)
3DRSSNet 3D representation Indoor environment CNN 4.4 - 5.5
Krijestorac et. al. [13] 3D representation Ray tracing software CNN 6.95 - 11.11
Cheng et. al. [1] 3D vector representation Indoor environment ANN 15.52-16.21
Raj et. al. [14] No Floorplan utilized Manually measured values ANN 5.37-5.65
TABLE I: Table comparing 3DRSSNet with other model implementations
The model’s performance and generalization property was
tested by implementing it on a new different floorplan. Fig.
2 shows that the model performance an average loss of 5.5
dbms, denoting good generalization property.
Fig. 2: Loss Value of model tested on different floor plan.
An implementation using a 2D floor plan was also tested
to reinforce the need for using a 3D floor plan. The training
results show generally higher MAE loss, with a training loss
of around 30 dBms, and a validation loss of around 75 dBms.
These results concrete the need for a more profound, rich data
input and a matching model architecture to better learn from
the increased data. The use of a 3D floor plan along with a
deep networking model efficiently performs this task.
The implementation of the complete DL model
can be found on the GitHub repository at
https://github.com/YaqoobAnsari/Deep-Learning-of-Radio-
Link-Quality-in-Wireless-Networks-
VII. CON CL US IO N
This work has presented a deep learning algorithm that
can be used to predict signal propagation in a 3-D indoor
environment. Our algorithm uses 3D maps of the environment
to produce a stochastic prediction of received signal strength
values. It outperforms existing deep learning solutions for
signal strength prediction, with a substantial improvement
when the locations of transmitter and receiver are known.Our
algorithm has only been evaluated in a single-level indoor
setting with single access points. Our future work will involve
an application on multilevel structures with multiple access
points and the consideration of environmental noise in the
measurement.
ACK NOW LE DG ME NT
The research was funded by Carnegie Mellon University in
Qatar (CMU-Q) under the Qatar Student-Initiated Undergrad-
uate Research Program (QSIURP).
REF ER EN CE S
[1] Hong Cheng, Hyukjoon Lee, and Shengjie Ma. CNN-Based Indoor
Path Loss Modeling with Reconstruction of Input Images. 9th Inter-
national Conference on Information and Communication Technology
Convergence (ICTC) 2018, (1):605–610, 2018.
[2] Yaming Xu, Jianguo Zhou, and Peng Zhang. Rss-based source localiza-
tion when path-loss model parameters are unknown. IEEE communica-
tions letters, 18(6):1055–1058, 2014.
[3] Hemant Kumar Rath, Sumanth Timmadasari, Bighnaraj Panigrahi, and
Anantha Simha. Realistic indoor path loss modeling for regular wifi
operations in india. In Proc. of 2017 Twenty-third National Conference
on Communications (NCC), pages 1–6. IEEE, 2017.
[4] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.
Nature, 521(7553):436–444, 2015.
[5] Yi Feng, Linlan Liu, and Jian Shu. A link quality prediction method for
wireless sensor networks based on xgboost. IEEE Access, 7:155229–
155241, 2019.
[6] Xionghui Luo, Linlan Liu, Jian Shu, and Manar Al-Kali. Link Quality
Estimation Method for Wireless Sensor Networks Based on Stacked
Autoencoder. IEEE Access, 7:21572–21583, 2019.
[7] Aliye ¨
Ozge Kaya, Larry J. Greenstein, and Wade Trappe. Characterizing
indoor wireless channels via ray tracing combined with stochastic
modeling. IEEE Trans. Wirel. Commun., 8(8):4165–4175, 2009.
[8] Jun G.Yong Lee, Min Young Kang, and Seong Cheol Kim. Path Loss
Exponent Prediction for Outdoor Millimeter Wave Channels through
Deep Learning. In IEEE Wirel. Commun. Netw. Conf. WCNC, 2019.
[9] Chaoyun Zhang, Paul Patras, and Hamed Haddadi. Deep Learning in
Mobile and Wireless Networking: A Survey. IEEE Commun. Surv.
Tutorials, 21(3):2224–2287, 2019.
[10] Yves Teganya and Daniel Romero. Data-Driven Spectrum Cartography
via Deep Completion Autoencoders. In Proc. of the IEEE International
Conference on Communications (ICC), pages 1–7. IEEE, jun 2020.
[11] Ron Levie, Cagkan Yapar, Gitta Kutyniok, and Giuseppe Caire. Ra-
diounet: Fast radio map estimation with convolutional neural networks.
IEEE Transactions on Wireless Communications, pages 1–1, 2021.
[12] Leandro Car´
ısio Fernandes and Antonio Jos´
e Martins Soares. On the
use of image segmentation for propagation path loss prediction. In Proc.
of SBMO/IEEE MTT-S, pages 129–133, 2011.
[13] Enes Krijestorac, Samer Hanna, and Danijela Cabric. Spatial signal
strength prediction using 3d maps and deep learning. arXiv preprint
arXiv:2011.03597, 2020.
[14] Nibin Raj. Indoor rssi prediction using machine learning for wireless
networks. In 2021 International Conference on COMmunication Systems
& NETworkS (COMSNETS), pages 372–374. IEEE, 2021.
[15] T.K. Sarkar, Zhong Ji, Kyungjung Kim, A. Medouri, and M. Salazar-
Palma. A survey of various propagation models for mobile communi-
cation. IEEE Antennas and Propagation Magazine, 45(3):51–82, 2003.
[16] Ileana Popescu, Athanasios Kanstas, Evangelos Angelou, L Nafornita,
and Philip Constantinou. Applications of generalized rbf-nn for path
loss prediction. In The 13th IEEE intl. symposium on personal, indoor
and mobile radio communications, volume 1, pages 484–488, 2002.
[17] Joseph M Mom, Callistus O Mgbe, and Gabriel A Igwue. Application of
artificial neural network for path loss prediction in urban macrocellular
environment. Am. J. Eng. Res, 3(2):270–275, 2014.
